Pith. sign in

REVIEW 6 major objections 4 minor 2 cited by

DatasetAgent: A Novel Multi-Agent System for Auto-Constructing Datasets from Real-World Images

T0 review · 6 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A four-agent system automatically builds image datasets from real-world photos, and downstream vision models trained on them improve across classification, detection, and segmentation.

desk verdict A genuinely novel system paper whose evaluation doesn't yet isolate curation from added data volume. read the letter →

arxiv 2507.08648 v1 pith:CRXCR5OW submitted 2025-07-11 cs.CV cs.AI

Haoran Sun , Haoyu Bian , Shaoning Zeng , Yunbo Rao , Xu Xu , Lin Mei , Jianping Gou This is my paper
classification cs.CVcs.AI
keywords DatasetAgentmulti-agentsystemconstructionmultimodallargelanguagemodelsimageclassificationobjectdetectionsegmentationreal-worldimages
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

DatasetAgent is a proposed system that automatically builds image datasets from real-world images by coordinating four agents powered by multimodal large language models (MLLMs). The paper's central claim is that these automatically constructed datasets meet user-specified requirements and are high enough in quality that vision models trained on them perform consistently better than models trained on the original datasets. The authors demonstrate the claim in two settings: expanding existing datasets (CIFAR-10, STL-10, PASCAL VOC 2007, CamVid) and creating new datasets from scratch from web-collected images. Across classification, object detection, and segmentation, downstream accuracy improves, with average classification gains of 0.41–0.52 percentage points and mAP@0.5 gains of 2.95 points on the VOC expansion. If the claim holds, manual collection and annotation can be replaced with an automated pipeline that still produces data that improves model performance.

What carries the argument

The system's core mechanism is a four-agent orchestration loop. A Demand Analysis Agent turns the user requirement into a structured specification; an Image Process Agent collects images from the web, converts each candidate into a structured JSON description of semantic attributes and quality indicators, and invokes a Tool Package (cropping, resizing, color adjustment, normalization) to optimize accepted images; a Data Label Agent assigns labels and, for detection and segmentation, generates annotations and masks using vision-language models; and a Supervision Agent logs progress, diagnoses failures, and restarts from checkpoints. The load-bearing operation is MLLM-mediated semantic alignment: at every stage, a multimodal model judges whether an image matches the target class and quality bar, and these judgments determine which raw images enter the dataset.

What would settle it

Run a size-matched control: take the raw images collected for one expansion, train the same architecture on a random subset of the same size and labels that DatasetAgent would have kept, and compare accuracy; if the random subset matches the filtered one, the filtering and optimization steps are not the source of the reported improvement.

Watch

Extended reading notes

Core claim

The paper argues that DatasetAgent, a collaborative system of four agents built around multimodal large language models, can construct high-quality image datasets from real-world images with no human intervention beyond a brief natural-language description. The authors claim that the datasets it builds are high quality on six evaluation metrics (class balance, structural similarity, annotation label reliability, data source entropy, sample diversity, and distribution consistency) and that training downstream vision models on them consistently improves performance. In the expansion experiments, adding DatasetAgent-curated images to CIFAR-10 and STL-10 improves classification accuracy by 0.41–0.52 percentage points on average across eight models, and expanding PASCAL VOC 2007 raises mAP@0.5 by up to 4.7 points for YOLOv8. In the from-scratch experiments, two new ten-class datasets reach average classification accuracy of 98.90% and 96.77%, and a newly built detection/segmentation dataset reaches quality metrics comparable to expanded standard datasets. The conclusion the paper draws is that fully automated construction of image datasets from real-world images is feasible and beneficial for downstream learning.

Load-bearing premise

The whole pipeline assumes the multimodal model's judgments about which images are relevant, clean, and correctly labeled are trustworthy enough that errors do not poison the dataset.

Editorial extensions

If this is right

  • Fully automated dataset construction from real-world images becomes practical, with no human annotation step along the collection, cleaning, and labeling pipeline.
  • Expanding a standard classification or detection dataset with DatasetAgent-curated images raises downstream accuracy, so dataset owners can grow their data without manual labor.
  • Creating new datasets from scratch is feasible for image classification, object detection, and segmentation, including pixel-level and panoptic masks.
  • Real-world imagery plus automated curation can replace or reduce reliance on synthetic image generation for training-data production.
  • The reported quality metrics provide a reusable audit template for evaluating whether an auto-constructed dataset is balanced, diverse, and reliably labeled.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: because the expansion comparisons pit a larger auto-curated dataset against a smaller original one, part of the accuracy gain could simply be more training data; a size-matched random-selection control would isolate what the curation itself contributes.
  • Editorial inference: the dataset-quality metrics are produced inside the same MLLM-driven loop, so an independent human audit on a sample of images would test whether those metrics actually track downstream model performance.
  • Editorial inference: the system's design is not limited to the benchmark categories tested, so extending it to medical, remote sensing, or industrial imagery is a plausible next application, but the paper provides no evidence for those domains.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

6 major / 4 minor

Summary. The paper proposes DatasetAgent, a multi-agent system with Demand Analysis, Image Process, Data Label, and Supervision agents, plus a tool package, intended to automatically construct image datasets from real-world images. The system is evaluated in two modes: expanding existing datasets (CIFAR-10, STL-10, PASCAL VOC2007, CamVid) and 'creating' new datasets, with downstream models trained for classification, object detection, and segmentation. The central claims are that DatasetAgent constructs high-quality datasets automatically and that these constructed datasets yield consistent improvements in downstream task performance, with average classification accuracy up to 98.90%. The paper also introduces a battery of quality metrics (CBI, SSIM, ALR, DSE, SDI, DDC, and task-specific ones) to support the quality claim.

Significance. The problem is timely and important: fully automated construction of real-world image datasets would reduce human annotation cost and avoid the domain gap of synthetic data. The paper deserves credit for testing across three task families with a broad set of models and for explicitly acknowledging limitations in Appendix D. If the central claims were established, this would be a useful systems contribution. However, the evidence as presented does not support those claims: the 'two orders of magnitude' expansion is contradicted by the paper's own Table 2, the downstream comparisons do not control for increased training volume, several quality metrics are self-referential or undefined in their reference distributions, and the 'from scratch' claim is contradicted by the paper's own text. The internal contradictions and the absence of a proper control condition mean the current contribution is not yet substantiated.

major comments (6)
  1. [Section 4.1, Table 2] The claim that the agent expands CIFAR-10 and STL-10 'by two orders of magnitude' is not supported by the reported numbers. Table 2 shows final sizes of 70,188 and 75,101 for CIFAR-10 (from 60,000), and 25,982 and 29,832 for STL-10 (from 13,000). These are increases of approximately 1.2x and 2x, not 100x; the added image counts are 10,188/15,101 and 12,982/16,832, respectively. The sentence in Section 4.3 about 'accuracy enhancements by two orders of magnitude' is also unsupported, since all reported accuracy changes are below 2.5 percentage points. This is a load-bearing quantitative claim and must be corrected.
  2. [Section 4.1, Tables 3, 5, and 7] The downstream evaluations do not isolate the effect of DatasetAgent's curation from the effect of increased training set size. Each 'New' condition adds thousands of images (e.g., CIFAR-10 +10,188/+15,101; VOC2007 +5,231/+6,179; CamVid +536/+1,088), and no control condition trains on an equally sized set of randomly sampled or unfiltered images. The reported gains of 0.41–0.52% in classification and a few mAP/MIoU points are entirely consistent with a volume effect. Moreover, the claim of 'consistent improvement' is directly contradicted by GoogLeNet, which drops from 89.69% to 87.59% on CIFAR-10(1) and from 89.69% to 88.31% on CIFAR-10(2). Without a same-volume uncurated baseline and multiple runs, the manuscript does not establish that curation, rather than added data, causes the observed improvements.
  3. [Section 4.2, Listing 1, Appendix D] The quality evaluation is partly circular and partly undocumented. The 'semantic_alignment', 'quality_risks', and 'decision' fields in Listing 1 are generated by the same MLLM-based pipeline that selects and filters images in Section 3.3, and the SDI and DDC metrics are computed from feature representations of that pipeline. These therefore reflect the system's own judgments rather than an independent ground truth. ALR is described only as 'measured by manual inspection' with no protocol, inspector count, or inter-rater reliability; BQI and ACS similarly lack procedural detail. The KL divergence in DDC is written as KL(P||Q) without defining P and Q. Appendix D further concedes that the pipeline is limited by large-model performance and has high rejection rates for complex or uncommon objects, which tempers the quality guarantees. The metrics need external validation and complete definitions.
  4. [Sections 4.1, 4.3, and 5] The 'from scratch' claim is internally inconsistent. Section 4.1 describes 'building entirely new datasets from scratch,' but Section 4.3 states that Dataset-Ours(1) and Dataset-Ours(2) are 'two curated ten-class datasets from ImageNet,' with Table 8 labeling them ImageNet-10. The detection and segmentation datasets are likewise 'curated from multiple existing large-scale datasets' (Section 4.1) rather than collected from raw real-world sources. In addition, Section 5 states that 'Currently, DatasetAgent can only construct image classification datasets for image classification tasks,' which directly contradicts the detection and segmentation experiments in Sections 4.4 and 4.5. These contradictions bear on the core claim of automated multi-task dataset construction and must be resolved.
  5. [Tables 3, 5, and 7; Section 4.1] No statistical reliability is reported. All downstream results appear to be single runs with no seeds, standard deviations, or significance tests. Several classification differences are very small (e.g., VGG-11 on CIFAR-10(1): 86.59% to 86.81%; DenseNet-121 on CIFAR-10(1): 85.81% to 85.89%) and are within typical run-to-run variability for these models trained from scratch. The manuscript states that 'identical hyperparameters and training cycles' were used but does not state the number of runs or random seeds. At least three to five seeds with confidence intervals are needed to support the 'consistent improvement' claim.
  6. [Section 3 and Appendix B] The reproducibility of the system is insufficiently documented. No code, prompts, model version identifiers, or tool-package API details are provided, even though the paper claims a fully automated and general system. In addition, Table 8 reports 'Score(%)' for different MLLMs, but this score is never defined in Section 4.2 or elsewhere; Appendix B also refers to 'scores of DatasetAgents' without explaining what is being scored. Without these definitions and artifacts, the system cannot be reconstructed or compared against future work.
minor comments (4)
  1. [Throughout] There are several typographical and naming inconsistencies: 'LLaV A' should be 'LLaVA', 'Grounding Dino' should be 'Grounding DINO', 'Ori' and 'Original' are used interchangeably, and 'Oct 5' should be a proper date. In the version I reviewed, Figures 5–8 are referenced in the text but not present in the manuscript body; please verify that all figures are included.
  2. [Section 4.2] The definitions of DSE and DDC are incomplete: Data Source Entropy needs a precise definition of what constitutes a 'source' for each dataset, and the KL divergence in DDC must specify which distributions are P and Q and how they are estimated from the data.
  3. [Section 4.1 and Table 2] The table formatting is confusing because multiple numbers appear in a single cell without clear sub-labels. It would be clearer to separate 'Original', 'Collected', and 'New' into distinct rows or columns, and to state explicitly that 'New' is the final total dataset size, not the number of added images.
  4. [Appendix D] The first listed limitation states that the method has 'higher operational costs and time' than traditional manual annotation, but no quantitative comparison is provided; the claim that it is 'still more efficient and cost-effective' needs supporting measurements or a reference.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: downstream accuracy on standard benchmarks is an external signal, and the construction pipeline is not fitting to that signal.

full rationale

DatasetAgent's central claim is that its constructed datasets improve downstream models. The evidence for this claim is measured test accuracy on standard benchmarks (CIFAR-10, STL-10, VOC2007, CamVid), which is external to the pipeline and not generated by the DatasetAgent itself. No component of the system is fitted to the test sets, and no reported prediction is defined in terms of the evaluation outcome. The quality metrics in Tables 2, 4, and 6 are either standard image statistics (SSIM, CBI, DSE) or human inspection (ALR, BQI, ACS); even if SDI and DDC are computed from model-derived features, they are descriptive quality indicators rather than the causal evidence for downstream improvement. Appendix D explicitly concedes that the system is limited by the underlying large models, which is a limitation statement rather than a self-certifying loop. The principal weakness is that the expansion experiments add both curation and training volume simultaneously, with no equal-volume uncurated control, so the causal contribution of curation is underdetermined; this is a confounding/control limitation, not a reduction of the claim to its inputs by construction. The only apparent self-citation (AUNet, reference [66]) is used as one of several segmentation models in the evaluation suite and is not load-bearing. Accordingly, no specific circular step can be identified from the text, and the appropriate score is 0.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The system introduces no new physical or scientific entities. The four agents are software roles, not new conceptual objects. The main free parameters are the hand-chosen thresholds and the internal MLLM scoring cutoff, which affect the composition of every constructed dataset and therefore every reported quality metric and downstream accuracy.

free parameters (3)
  • annotation confidence threshold = 0.5
    Section 3.4 states a minimum confidence score of 0.5 for inclusion of detected objects in detection/segmentation datasets. This value is hand-chosen and directly affects dataset composition and downstream metrics.
  • dataset quality thresholds = CBI<0.1, SSIM>0.9, DDC<0.1, ACS>0.85, PCB>0.8
    Section 4.2 defines acceptance thresholds for declaring datasets high-quality. These are arbitrary standards set by the authors, not derived from external benchmarks.
  • MLLM semantic alignment cutoff = not specified numerically
    Listing 1 shows a similarity score of 0.931 and a total risk score of 0.07 leading to 'qualified: true', implying a decision threshold inside the MLLM. The exact cutoff is not given and is effectively a free parameter of the curation pipeline.
assumptions (3)
  • domain assumption Real-world images are more valuable than artificially generated images for training datasets.
    Stated in the Abstract and Section 1 as 'common knowledge' and 'obviously more valuable'. The paper does not empirically test this claim, yet it motivates the entire system.
  • domain assumption Multimodal LLM semantic analysis is sufficiently accurate to guide image selection and annotation.
    Sections 3.3 and 3.4 rely on MLLM outputs for semantic alignment, quality scoring, and label generation. The paper's Appendix D concedes this assumption is fragile for complex scenes and uncommon objects.
  • domain assumption Identical training procedures across compared models and datasets provide a fair evaluation.
    Section 4.1 claims identical hyperparameters, Adam optimizer, and pre-trained models, but no details of hyperparameter values or training schedules are given, so the assumption cannot be verified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DatasetAgent: A Novel Multi-Agent System for Auto-Constructing Datasets from Real-World Images." pith.science (2026). https://pith.science/paper/CRXCR5OW

@misc{pith2026250708648,
  author       = {Pith},
  title        = {Pith review of: DatasetAgent: A Novel Multi-Agent System for Auto-Constructing Datasets from Real-World Images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CRXCR5OW}},
  note         = {Machine review of arXiv:2507.08648}
}
read the original abstract

Common knowledge indicates that the process of constructing image datasets usually depends on the time-intensive and inefficient method of manual collection and annotation. Large models offer a solution via data generation. Nonetheless, real-world data are obviously more valuable comparing to artificially intelligence generated data, particularly in constructing image datasets. For this reason, we propose a novel method for auto-constructing datasets from real-world images by a multiagent collaborative system, named as DatasetAgent. By coordinating four different agents equipped with Multi-modal Large Language Models (MLLMs), as well as a tool package for image optimization, DatasetAgent is able to construct high-quality image datasets according to user-specified requirements. In particular, two types of experiments are conducted, including expanding existing datasets and creating new ones from scratch, on a variety of open-source datasets. In both cases, multiple image datasets constructed by DatasetAgent are used to train various vision models for image classification, object detection, and image segmentation.

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. An Uncertainty-Driven Adaptive Self-Alignment Framework for Large Language Models

    cs.AI 2025-07 conditional novelty 5.0 of 10

    A training pipeline that scores a model's own responses for semantic, factual, and safety uncertainty, builds preference pairs from those scores, and trains in three difficulty stages improves reported alignment score...

  2. A Novel Self-Evolution Framework for Large Language Models

    cs.CL 2025-07 reject novelty 4.0 of 10

    A dual-phase framework that uses a Censor satisfaction scorer to expand training data and then applies SFT plus frequency-weighted DPO, reporting benchmark gains over SFT, PO, and memory baselines.

Reference graph

Works this paper leans on

82 extracted references · 49 canonical work pages · cited by 2 Pith papers

  1. [1]

    Abdullin, D

    Y . Abdullin, D. Molla-Aliod, B. Ofoghi, J. Yearwood, and Q. Li. Syn- thetic dialogue dataset generation using llm agents.arXiv preprint arXiv:2401.17461, 2024

    arXiv 2024

  2. [2]

    Abras, S

    S. Abras, S. Ploix, S. Pesty, and M. Jacomino. A multi-agent home automation system for power management. InInformatics in control automation and robotics: selected papers from the international con- ference on informatics in control automation and robotics 2006, pages 59–68. Springer, 2008

    2006

  3. [3]

    F. F. Alkhalid, A. Q. Albayati, and A. A. Alhammad. Expansion dataset covid-19 chest x-ray using data augmentation and histogram equaliza- tion.International Journal of Electrical and Computer Engineering, 12 (2):1904–1909, 2022

    1904

  4. [4]

    Bansal, M

    M. Bansal, M. Kumar, M. Sachdeva, and A. Mittal. Transfer learning for image classification using vgg19: Caltech-101 image data set.Journal of ambient intelligence and humanized computing, pages 1–12, 2023

    work page 2023

  5. [5]

    A. Y . Barrera-Animas and J. M. D. Delgado. Generating real-world-like labelled synthetic datasets for construction site applications.Automa- tion in Construction, 151:104850, 2023

    work page 2023

  6. [6]

    Bose and V

    R. Bose and V . Sugumaran. Application of intelligent agent technology for managerial data analysis and mining.ACM SIGMIS Database: the DATABASE for Advances in Information Systems, 30(1):77–94, 1999

    work page 1999

  7. [7]

    Bousmalis, G

    K. Bousmalis, G. Vezzani, D. Rao, C. Devin, A. X. Lee, M. Bauza, T. Davchev, Y . Zhou, A. Gupta, A. Raju, et al. Robocat: A self- improving foundation agent for robotic manipulation.arXiv preprint arXiv:2306.11706, 2023

    arXiv 2023

  8. [8]

    G. J. Brostow, J. Shotton, J. Fauqueur, and R. Cipolla. Segmentation and recognition using structure from motion point clouds. InCom- puter vision–ECCV 2008: 10th European conference on computer vi- sion, marseille, France, October 12-18, 2008, proceedings, part i 10, pages 44–57. Springer, 2008

    work page 2008

Show all 82 references
  1. [9]

    Y . Cao, S. Li, Y . Liu, Z. Yan, Y . Dai, P. S. Yu, and L. Sun. A compre- hensive survey of ai-generated content (aigc): A history of generative ai from gan to chatgpt.arXiv preprint arXiv:2303.04226, 2023

    2023 arXiv

  2. [10]

    Chang, X

    Y . Chang, X. Wang, J. Wang, Y . Wu, L. Yang, K. Zhu, H. Chen, X. Yi, C. Wang, Y . Wang, et al. A survey on evaluation of large language models.ACM Transactions on Intelligent Systems and Technology, 15 (3):1–45, 2024

    2024

  3. [11]

    H. Chen, K. Sun, Z. Tian, C. Shen, Y . Huang, and Y . Yan. Blendmask: Top-down meets bottom-up for instance segmentation. InProceedings of the IEEE/CVF conference on computer vision and pattern recogni- tion, pages 8573–8581, 2020

    2020

  4. [12]

    L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs.IEEE transactions on pat- tern analysis and machine intelligence, 40(4):834–848, 2017

    2017

  5. [13]

    Cheng, A

    B. Cheng, A. Schwing, and A. Kirillov. Per-pixel classification is not all you need for semantic segmentation.Advances in neural information processing systems, 34:17864–17875, 2021

    2021

  6. [14]

    Coates, A

    A. Coates, A. Ng, and H. Lee. An analysis of single-layer networks in unsupervised feature learning. InProceedings of the fourteenth inter- national conference on artificial intelligence and statistics, pages 215–

  7. [15]

    Cordts, M

    M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. Enzweiler, R. Benen- son, U. Franke, S. Roth, and B. Schiele. The cityscapes dataset for semantic urban scene understanding. InProceedings of the IEEE con- ference on computer vision and pattern recognition, pages 3213–3223, 2016

    2016

  8. [16]

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. Imagenet: A large-scale hierarchical image database. In2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009

    2009

  9. [17]

    A. M. DiGiorgio and J. M. Ehrenfeld. Artificial intelligence in medicine & chatgpt: de-tether the physician.Journal of Medical Systems, 47(1): 32, 2023

    2023

  10. [18]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby. An image is worth 16x16 words: Trans- formers for image recognition at scale, 2021. URL https://arxiv.org/abs/ 2010.11929

    2021 arXiv

  11. [19]

    Everingham, L

    M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2007 (VOC2007) Results. http://www.pascal- network.org/challenges/VOC/voc2007/workshop/index.html

    2007

  12. [20]

    Everingham, L

    M. Everingham, L. Van Gool, C. K. Williams, J. Winn, and A. Zisser- man. The pascal visual object classes (voc) challenge.International journal of computer vision, 88:303–338, 2010

    2010

  13. [21]

    Everingham, S

    M. Everingham, S. A. Eslami, L. Van Gool, C. K. Williams, J. Winn, and A. Zisserman. The pascal visual object classes challenge: A retro- spective.International journal of computer vision, 111:98–136, 2015

    2015

  14. [22]

    Gil-Martín, C

    M. Gil-Martín, C. Luna-Jiménez, S. Esteban-Romero, M. Estecha- Garitagoitia, F. Fernández-Martínez, and L. F. D’Haro. A dataset of synthetic art dialogues with chatgpt.Scientific Data, 11(1):825, 2024

    2024

  15. [23]

    Girshick

    R. Girshick. Fast r-cnn. InProceedings of the IEEE international con- ference on computer vision, pages 1440–1448, 2015

    2015

  16. [24]

    Girshick, J

    R. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich feature hier- archies for accurate object detection and semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 580–587, 2014

    2014

  17. [25]

    R. Gong, Q. Huang, X. Ma, H. V o, Z. Durante, Y . Noda, Z. Zheng, S.- C. Zhu, D. Terzopoulos, L. Fei-Fei, et al. Mindagent: Emergent gaming interaction.arXiv preprint arXiv:2309.09971, 2023

    2023 arXiv

  18. [26]

    K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition, 2015. URL https://arxiv.org/abs/1512.03385

    2015 arXiv

  19. [27]

    K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

    2016

  20. [28]

    K. He, G. Gkioxari, P. Dollár, and R. Girshick. Mask r-cnn. InProceed- ings of the IEEE international conference on computer vision, pages 2961–2969, 2017

    2017

  21. [29]

    Huang, Z

    G. Huang, Z. Liu, L. van der Maaten, and K. Q. Weinberger. Densely connected convolutional networks, 2018. URL https://arxiv.org/abs/ 1608.06993

    2018 arXiv

  22. [30]

    Isazawa and J

    T. Isazawa and J. M. Cole. automated construction of a photocataly- sis dataset for water-splitting applications.Scientific Data, 10(1):651, 2023

    2023

  23. [31]

    Jiang, Z

    H. Jiang, Z. Diao, T. Shi, Y . Zhou, F. Wang, W. Hu, X. Zhu, S. Luo, G. Tong, and Y .-D. Yao. A review of deep learning-based multiple- lesion recognition from medical images: classification, detection and segmentation.Computers in Biology and Medicine, 157:106726, 2023

    2023

  24. [32]

    L. Jiao, F. Zhang, F. Liu, S. Yang, L. Li, Z. Feng, and R. Qu. A survey of deep learning-based object detection.IEEE access, 7:128837–128868, 2019

    2019

  25. [33]

    Kawano and K

    Y . Kawano and K. Yanai. Automatic expansion of a food image dataset leveraging existing categories with domain adaptation. InComputer Vision-ECCV 2014 Workshops: Zurich, Switzerland, September 6-7 and 12, 2014, Proceedings, Part III 13, pages 3–17. Springer, 2015

    2014

  26. [34]

    M. R. King. The future of ai in medicine: a perspective from a chatbot. Annals of Biomedical Engineering, 51(2):291–295, 2023

    2023

  27. [35]

    Krizhevsky, G

    A. Krizhevsky, G. Hinton, et al. Learning multiple layers of features from tiny images. 2009

    2009

  28. [36]

    Kurt Pehlivano ˘glu, R

    M. Kurt Pehlivano ˘glu, R. T. Gobosho, M. A. Syakura, V . Shanmu- ganathan, and L. de-la Fuente-Valentín. Comparative analysis of para- phrasing performance of chatgpt, gpt-3, and t5 language models using a new chatgpt generated dataset: Paragpt.Expert Systems, page e13699, 2024

    2024

  29. [37]

    Lee and J

    Y . Lee and J. Park. Centermask: Real-time anchor-free instance seg- mentation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13906–13915, 2020

    2020

  30. [38]

    C. Li, Z. Gan, Z. Yang, J. Yang, L. Li, L. Wang, J. Gao, et al. Mul- timodal foundation models: From specialists to general-purpose assis- tants.Foundations and Trends® in Computer Graphics and Vision, 16 (1-2):1–214, 2024

    2024

  31. [39]

    Y . Li, Q. Li, L. Cui, W. Bi, Z. Wang, L. Wang, L. Yang, S. Shi, and Y . Zhang. Mage: Machine-generated text detection in the wild. InPro- ceedings of the 62nd Annual Meeting of the Association for Computa- tional Linguistics (Volume 1: Long Papers), pages 36–53, 2024

    2024

  32. [40]

    Y . Li, H. Wen, W. Wang, X. Li, Y . Yuan, G. Liu, J. Liu, W. Xu, X. Wang, Y . Sun, et al. Personal llm agents: Insights and survey about the capa- bility, efficiency and security.arXiv preprint arXiv:2401.05459, 2024

    2024 arXiv

  33. [41]

    Z. Li, W. Wang, E. Xie, Z. Yu, A. Anandkumar, J. M. Alvarez, P. Luo, and T. Lu. Panoptic segformer: Delving deeper into panoptic segmen- tation with transformers. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1280–1289, 2022

    2022

  34. [42]

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick. Microsoft coco: Common objects in context. InComputer vision–ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13, pages 740–755. ...

    2014

  35. [43]

    T.-Y . Lin, P. Goyal, R. Girshick, K. He, and P. Dollár. Focal loss for dense object detection. InProceedings of the IEEE international con- ference on computer vision, pages 2980–2988, 2017

    2017

  36. [44]

    H. Liu, C. Li, Q. Wu, and Y . J. Lee. Visual instruction tuning, 2023

    2023

  37. [45]

    X. Liu, H. Yu, H. Zhang, Y . Xu, X. Lei, H. Lai, Y . Gu, H. Ding, K. Men, K. Yang, et al. Agentbench: Evaluating llms as agents.arXiv preprint arXiv:2308.03688, 2023

    2023 arXiv

  38. [46]

    Y . Liu, C. Zhang, and X. Dong. A survey of real-time surface defect in- spection methods based on deep learning.Artificial Intelligence Review, 56(10):12131–12170, 2023

    2023

  39. [47]

    Z. Liu, W. Yao, J. Zhang, L. Yang, Z. Liu, J. Tan, P. K. Choubey, T. Lan, J. Wu, H. Wang, et al. Agentlite: A lightweight library for building and advancing task-oriented llm agent system.arXiv preprint arXiv:2402.15538, 2024

    2024 arXiv

  40. [48]

    J. Long, E. Shelhamer, and T. Darrell. Fully convolutional networks for semantic segmentation. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 3431–3440, 2015

    2015

  41. [49]

    Minaee, Y

    S. Minaee, Y . Boykov, F. Porikli, A. Plaza, N. Kehtarnavaz, and D. Ter- zopoulos. Image segmentation using deep learning: A survey.IEEE transactions on pattern analysis and machine intelligence, 44(7):3523– 3542, 2021

    2021

  42. [50]

    Minaee, T

    S. Minaee, T. Mikolov, N. Nikzad, M. Chenaghlu, R. Socher, X. Am- atriain, and J. Gao. Large language models: A survey.arXiv preprint arXiv:2402.06196, 2024

    2024 arXiv

  43. [51]

    Neuhold, T

    G. Neuhold, T. Ollmann, S. Rota Bulo, and P. Kontschieder. The map- illary vistas dataset for semantic understanding of street scenes. In Proceedings of the IEEE international conference on computer vision, pages 4990–4999, 2017

    2017

  44. [53]

    Redmon and A

    J. Redmon and A. Farhadi. Yolov3: An incremental improvement.arXiv preprint arXiv:1804.02767, 2018

    2018 arXiv

  45. [54]

    Redmon, S

    J. Redmon, S. Divvala, R. Girshick, and A. Farhadi. You only look once: Unified, real-time object detection. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 779–788, 2016

    2016

  46. [55]

    T. Ren, Q. Jiang, S. Liu, Z. Zeng, W. Liu, H. Gao, H. Huang, Z. Ma, X. Jiang, Y . Chen, et al. Grounding dino 1.5: Advance the" edge" of open-set object detection.arXiv preprint arXiv:2405.10300, 2024

    2024 arXiv

  47. [56]

    T. Ren, S. Liu, A. Zeng, J. Lin, K. Li, H. Cao, J. Chen, X. Huang, Y . Chen, F. Yan, Z. Zeng, H. Zhang, F. Li, J. Yang, H. Li, Q. Jiang, and L. Zhang. Grounded sam: Assembling open-world models for diverse visual tasks, 2024

    2024

  48. [57]

    Ronneberger, P

    O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolutional networks for biomedical image segmentation. InMedical image computing and computer-assisted intervention–MICCAI 2015: 18th international con- ference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pa...

    2015

  49. [58]

    M. Roser. The brief history of artificial intelligence: the world has changed fast—what might be next?Our world in data, 2024

    2024

  50. [59]

    B. C. Russell, A. Torralba, K. P. Murphy, and W. T. Freeman. Labelme: a database and web-based tool for image annotation.International jour- nal of computer vision, 77:157–173, 2008

    2008

  51. [60]

    Simonyan and A

    K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition.arXiv preprint arXiv:1409.1556, 2014

    2014 arXiv

  52. [61]

    Simonyan and A

    K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition, 2015. URL https://arxiv.org/abs/1409. 1556

    2015

  53. [62]

    M. M. Soltani, Z. Zhu, and A. Hammad. Automated annotation for vi- sual recognition of construction resources using synthetic images.Au- tomation in Construction, 62:14–23, 2016

    2016

  54. [63]

    J. Song, S. Gao, Y . Zhu, and C. Ma. A survey of remote sensing image classification based on cnns.Big earth data, 3(3):232–254, 2019

    2019

  55. [64]

    Srivastava, G

    N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhut- dinov. Dropout: a simple way to prevent neural networks from over- fitting.The journal of machine learning research, 15(1):1929–1958, 2014

    1929

  56. [65]

    Stefanics and M

    D. Stefanics and M. Fox. Coco annotator: Web-based image segmen- tation tool for object detection, localization, and keypoints.ACM SIG- Multimedia Records, 13(3):1–1, 2022

    2022

  57. [66]

    H. Sun, C. Li, B. Liu, Z. Liu, M. Wang, H. Zheng, D. D. Feng, and S. Wang. Aunet: attention-guided dense-upsampling networks for breast mass segmentation in whole mammograms.Physics in Medicine & Biology, 65(5):055005, 2020

    2020

  58. [67]

    Szegedy, W

    C. Szegedy, W. Liu, Y . Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V . Vanhoucke, and A. Rabinovich. Going deeper with convolutions,

  59. [68]

    Talebirad and A

    Y . Talebirad and A. Nadiri. Multi-agent collaboration: Harnessing the power of intelligent llm agents.arXiv preprint arXiv:2306.03314, 2023

    2023 arXiv

  60. [69]

    Tan and Q

    M. Tan and Q. V . Le. Efficientnet: Rethinking model scaling for convo- lutional neural networks, 2020. URL https://arxiv.org/abs/1905.11946

    2020 arXiv

  61. [70]

    Varghese and M

    R. Varghese and M. Sambath. Yolov8: A novel object detection algo- rithm with enhanced performance and robustness. In2024 International Conference on Advances in Data Engineering and Intelligent Comput- ing Systems (ADICS), pages 1–6. IEEE, 2024

    2024

  62. [71]

    L. Wang, C. Ma, X. Feng, Z. Zhang, H. Yang, J. Zhang, Z. Chen, J. Tang, X. Chen, Y . Lin, et al. A survey on large language model based autonomous agents.Frontiers of Computer Science, 18(6):186345, 2024

    2024

  63. [72]

    W. Wang, Y . Xu, C. Du, Y . Chen, Y . Wang, and H. Wen. Data set and evaluation of automated construction of financial knowledge graph. Data Intelligence, 3(3):418–443, 2021

    2021

  64. [73]

    Q. Wu, G. Bansal, J. Zhang, Y . Wu, S. Zhang, E. Zhu, B. Li, L. Jiang, X. Zhang, and C. Wang. Autogen: Enabling next-gen llm applications via multi-agent conversation framework.arXiv preprint arXiv:2308.08155, 2023

    2023 arXiv

  65. [74]

    S. Wu, H. Fei, L. Qu, W. Ji, and T.-S. Chua. Next-gpt: Any-to-any multimodal llm.arXiv preprint arXiv:2309.05519, 2023

    2023 arXiv

  66. [75]

    T. Wu, S. He, J. Liu, S. Sun, K. Liu, Q.-L. Han, and Y . Tang. A brief overview of chatgpt: The history, status quo and potential future devel- opment.IEEE/CAA Journal of Automatica Sinica, 10(5):1122–1136, 2023

    2023

  67. [76]

    Z. Xi, W. Chen, X. Guo, W. He, Y . Ding, B. Hong, M. Zhang, J. Wang, S. Jin, E. Zhou, et al. The rise and potential of large language model based agents: A survey.arXiv preprint arXiv:2309.07864, 2023

    2023 arXiv

  68. [77]

    Y . Yao, J. Zhang, F. Shen, X. Hua, J. Xu, and Z. Tang. Exploiting web images for dataset construction: A domain robust approach.IEEE Transactions on Multimedia, 19(8):1771–1784, 2017. doi: 10.1109/ TMM.2017.2684626

    2017

  69. [78]

    Yurtsever, J

    E. Yurtsever, J. Lambert, A. Carballo, and K. Takeda. A survey of au- tonomous driving: Common practices and emerging technologies.IEEE access, 8:58443–58469, 2020

    2020

  70. [79]

    A. Zeng, M. Liu, R. Lu, B. Wang, X. Liu, Y . Dong, and J. Tang. Agent- tuning: Enabling generalized agent abilities for llms.arXiv preprint arXiv:2310.12823, 2023

    2023 arXiv

  71. [80]

    Zhang, Y

    J. Zhang, Y . Hou, R. Xie, W. Sun, J. McAuley, W. X. Zhao, L. Lin, and J.-R. Wen. Agentcf: Collaborative learning with autonomous language agents for recommender systems. InProceedings of the ACM on Web Conference 2024, pages 3679–3689, 2024

    2024

  72. [223]

    JMLR Workshop and Conference Proceedings, 2011

    2011

  73. [2014]

    URL https://arxiv.org/abs/1409.4842

    arXiv

  74. [2024]

    URL https://arxiv.org/abs/2408.00714

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.